Robust Explanations Through Uncertainty Decomposition: A Path to Trustworthier AI

Arthur Hoarau; Chenrui Zhu; Louenas Bounia; S\'ebastien Destercke; Vu Linh Nguyen

REVIEW 5 major objections 7 minor 21 references

Robust Explanations Through Uncertainty Decomposition: A Path to Trustworthier AI

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that the type of prediction uncertainty tells you which explanation to offer: counterfactuals for high aleatoric uncertainty, feature importance for low, and an 'insufficient training' rejection for high epistemic…

desk verdict A plausible routing protocol for explanation choice, but the headline counterfactual correlation may be baked into the shared KNN machinery; worth refereeing with revision. read the letter →

arxiv 2507.12913 v1 pith:FJE4TCFQ submitted 2025-07-17 cs.LG

Chenrui Zhu , Louenas Bounia , Vu Linh Nguyen , Sébastien Destercke , Arthur Hoarau This is my paper

classification cs.LG

keywords uncertaintydecompositionaleatoricepistemiccounterfactualexplanationsfeatureimportanceSHAPexplanationrobustnessrejectoption

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that uncertainty decomposition can be used as a practical selector among explanation methods. The central proposal is to split prediction uncertainty into aleatoric (data noise/class overlap) and epistemic (lack of training data) parts; high epistemic uncertainty means no post-hoc explanation is trustworthy, and the uncertainty itself should be shown as an explanation: the model lacks training. For the remaining points, high aleatoric uncertainty points to counterfactual explanations, because class-confused regions tend to yield close, attainable counterfactuals, while low aleatoric uncertainty points to feature-importance explanations such as SHAP, which are more locally stable where classes are separated. If the correlations reported hold, explanation systems become context-aware rather than one-size-fits-all.

What carries the argument

The load-bearing objects are two uncertainty estimates derived from a Dempster-Shafer belief function on the K-nearest neighbors: epistemic uncertainty EUbel (non-specificity, Equation 12) and aleatoric uncertainty AUbel (discord, Equation 13). Around these sit the counterfactual defined as the nearest training point with a different predicted label (Equation 3), SHAP feature attributions (Equation 2), and two robustness metrics: discrete local Lipschitz continuity for feature-importance explanations (Equation 4) and Euclidean distance to the counterfactual as a measure of attainability (Equation 7). The argument runs by Spearman rank correlations between AUbel and these robustness scores, with the rejection rule driven by EUbel.

What would settle it

Compute AUbel and counterfactual distance on the same datasets but with different K values for the two computations (e.g., K=3 for the belief function and K=15 for the counterfactual search); if the negative Spearman correlation is robust, it should persist beyond a shared K=7 neighborhood. A second check is to construct a dataset with high class overlap but distant different-class neighbors, where the predicted negative correlation should vanish.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is a pair of empirical regularities plus a rejection rule. Across ten tabular datasets, the belief-based aleatoric uncertainty AUbel is significantly negatively correlated with counterfactual dissimilarity (Spearman's r roughly -0.45 to -0.89), meaning that in high-aleatoric regions the nearest different-class neighbor is closer and more attainable; and AUbel is positively correlated with SHAP un-robustness, meaning low-aleatoric regions give more locally stable feature-importance explanations. The paper further claims that epistemic uncertainty is itself a legitimate explanation: when it is high, the instance is out-of-distribution or under-represented, so the correct statement is that the model lacks training, and the usual explanation should be withheld. These findings hold across three uncertainty decomposition families and carry over to a deep learning MNIST example, where the most epistemically uncertain test image yields an unintelligible counterfactual that should be rejected.

Load-bearing premise

The central claim collapses if the negative AU-vs-counterfactual-attainability and positive AU-vs-SHAP-unrobustness correlations are largely artifacts of both quantities being computed from the same K-nearest-neighbor structure, rather than reflections of genuine data geometry.

Editorial extensions

If this is right

A deployed explanation system can use AUbel directly: when it is high, show the counterfactual; when it is low, show feature attributions.
High EUbel becomes a signal to withhold explanations and instead state that the model has insufficient evidence, turning explanation failure into a transparent outcome.
Local, density-based uncertainty estimates (belief-based and centroid-based) outperform global ensemble-entropy estimates as selectors, so more local uncertainty quantification is worth investing in.
The reject option gives a principled way to clean test sets before reporting SHAP values, since filtering the 30% most epistemically uncertain points often strengthens the AU-robustness correlation.
The protocol transfers from tabular classifiers to deep image classifiers, so the same uncertainty-decomposition guidance can be applied in deep learning settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves implicit that the same selector could be applied adaptively per instance at deployment time, for example a dashboard that switches explanation type based on AUbel.
Because AUbel and the counterfactual both come from the same K-nearest-neighbor structure, some of the observed correlation may be mechanical; varying K between the two computations would separate genuine data effects from shared-machinery artifacts.
A testable extension is to use the epistemic reject option as an acquisition criterion: instead of explaining uncertain points, query them for labels, since high EUbel marks regions where additional training data would most improve explanations.
A stricter test of the paper's claim would compare the recommended explanation type on the same low-AU and high-AU points against user-satisfaction or decision-quality measures, not just robustness metrics.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The one thing to know: this paper proposes a genuinely new protocol for choosing explanation type from uncertainty decomposition—counterfactuals when aleatoric uncertainty is high, SHAP when low, reject when epistemic is high—and backs it with a wide correlational study. The idea is plausible and useful, and the experimental breadth is real: ten UCI datasets, three UQ families, 100 runs for the counterfactual experiment, and mostly significant Spearman correlations.

The problem is that the paper's strongest pillar may be partly circular. The AUbel measure in Eq. (13) is the Discord over the K=7 neighborhood of a test point, while the counterfactual distance in Eq. (7) is the distance to the nearest different-class neighbor in the same training set, found with the same KNN classifier at K=7 (Appendix C). A test point that sits close to a different-class neighbor will mechanically tend to have high class conflict in that neighborhood, so the −0.83 correlations in Table 5 are at least partly a restatement of shared neighborhood structure. The independent estimators do not rescue it: AUens goes positive on Glass and Ionosphere, and AUcen is near zero on Ionosphere. The paper even concludes that "more local uncertainty quantification should be privileged," which effectively leans on the entangled estimator.

Other soft spots: the protocol is never benchmarked against a fixed explanation policy, so the correlations don't demonstrate that the routing improves user-facing outcomes; the SHAP experiment uses only 5 runs; the rejection evidence is anecdotal (one instance per setting); and code is promised only on acceptance. These are fixable.

That said, the paper is honest about some limits, the related work is well covered, and the central idea—using uncertainty to route explanations—is worth taking seriously. I'd send it to review, but flag the circularity question to the referees. If the authors re-run the counterfactual correlation with a distance computed outside the KNN structure (or with a separate held-out neighborhood), the claim would stand on firmer ground.

Referee Report

5 major / 7 minor

Summary. The paper proposes a two-part protocol for combining uncertainty quantification (UQ) with local explainability: (i) use aleatoric uncertainty (AU) to choose between counterfactual explanations (high AU) and feature-importance explanations (low AU), and (ii) use epistemic uncertainty (EU) as a reject option for explanations of predictions made in regions of missing training data. The empirical strategy is to measure Spearman correlations between AU estimates from three UQ methods (belief-based AUbel, centroid-based AUcen, ensemble-based AUens) and two explanation robustness metrics: the distance to the nearest counterfactual (Eq. 7) and the discrete local Lipschitz constant of SHAP explanations (Eq. 4). The authors report a significantly negative AU-vs-dissimilarity correlation (Table 1), a mostly positive AU-vs-un-robustness correlation (Table 2), and illustrative rejection curves for a Glass dataset and MNIST. They argue that uncertainty decomposition can serve as a practical guide to selecting more robust and attainable explanations.

Significance. The intersection of UQ and XAI is timely, and the proposed protocol is intuitive and potentially useful. The paper contributes a broad empirical study across 10 UCI datasets and MNIST, with three UQ families and a reproducible experimental setup (code is promised as supplementary material). The idea that epistemic uncertainty itself can be an explanation—'the model lacks training'—is a sensible and refreshing framing. However, the central empirical pillar—the negative AU-vs-counterfactual-dissimilarity correlation—is weakened by the fact that the belief-based estimator AUbel is constructed from the same K-nearest-neighbor geometry as the counterfactual definition. The independent estimators give mixed and sometimes sign-inconsistent results. Moreover, the paper never evaluates the full selection protocol against a fixed explanation policy, so the evidence as presented supports correlations but not the claimed benefit of the decision rule. If the tautology concern can be resolved and a protocol-level benchmark is added, the contribution would be valuable to the XAI community.

major comments (5)

[§3.1.2, Table 2] The headline negative correlation between AUbel and counterfactual dissimilarity (Table 1: -0.83, -0.89, -0.82 for several datasets) is at least partly a tautology arising from shared neighborhood structure. AUbel (Eq. 13) is computed from the belief function over the K=7 nearest neighbors of xt, and the counterfactual distance d (Eq. 7) is the Euclidean distance to the nearest training point with a different label (Eq. 3), which is exactly the kind of class conflict that increases the discord measure in Eq. (13). When a different-class neighbor lies inside the K=7 neighborhood, d is small and the local class conflict is high; when all K neighbors share a label, the nearest different-class point is farther away and the discord term is typically zero. Consequently, a strong negative Spearman correlation between AUbel and d is expected under a null model that has no genuine relationship between aleatoric uncertainty and counterfactual attainability. The independent estimators provide inconsistent evidence: AUens is positive on Glass (+0.15) and Ionosphere (+0.30), and AUcen is near zero on Ionosphere (-0.04). Since the paper explicitly concludes that 'more local uncertainty quantification should be privileged' (Section 3.1.1), the central empirical claim rests on the entangled measure. Please provide a control analysis that removes the shared-neighbor effect, e.g., by computing the AUbel vs. d correlation separately for test points whose nearest different-class neighbor is inside vs. outside the K=7 ball, or by a permutation test that randomizes labels within neighborhoods. Without such an analysis, the paper's main conclusion is not empirically supported.
[§3.1.2, Table 2] Table 2's heading says 'aleatoric uncertainty ... vs. robustness of feature-importance explanations', but the metric L(xt) in Eq. (4) is a Lipschitz constant, i.e., a measure of un-robustness (larger L means less robust). The text in §3.1.2 correctly speaks of 'un-robustness' and states that 'more aleatoric uncertainty leads to less robust SHAP explanations'. The positive correlations in Table 2 therefore mean high AU is associated with high un-robustness, which is the intended claim. However, the table and figure captions conflate the two directions, and a reader who takes the table at face value would draw the opposite conclusion. Please correct the labels in Table 2 and any related axes/captions to explicitly state that the correlation is with un-robustness (or equivalently, that positive correlation means AU is positively associated with the Lipschitz constant). Also clarify in the text why this direction is the one that supports the protocol.
[§3.1 and Appendix D] There is a direct inconsistency in the number of experimental runs. Section 3.1 states: 'Spearman's rank correlation are averaged over 100 runs and concern all decomposition methods.' Appendix D, however, states: 'Each experiment is repeated 5 times (due to SHAP expensive complexity).' The SHAP robustness results in Table 2 and Appendix D are thus based on only a few runs, not 100, and the main text's '100 runs' claim does not apply uniformly. Moreover, the p-values reported in Appendix D (e.g., Tables 9-12) are computed from a single test set and a single run, not from a distribution over runs. The number of runs should be stated consistently for each experiment, and the main text should avoid implying 100-run averages for the SHAP results.
[§3 overall] The paper proposes a decision rule: reject explanations when EU is high, select counterfactuals when AU is high, and select feature-importance when AU is low. Yet the experiments only report per-dataset correlations between AU and each explanation metric; they never evaluate the full rule against a fixed policy such as 'always counterfactual', 'always SHAP', or 'random selection'. A user following the protocol might not experience lower average dissimilarity or lower average Lipschitz constant if the correlations are dominated by extreme points or if the AU thresholds are poorly chosen. To substantiate the protocol's usefulness, please provide a benchmark: for each dataset, compute the average counterfactual dissimilarity (or SHAP un-robustness) under the AU-based selection rule and compare it with the average under a fixed policy and under random selection, with confidence intervals over the repeated runs. Without this comparative evaluation, the paper demonstrates relationships but not the value of the proposed protocol.
[§3.2 and §3.1.2] The reject option based on epistemic uncertainty is illustrated with only one extreme example on Glass and one on MNIST, and the rejection curves (Figures 5a, 6a) are not compared against any baseline (e.g., random rejection or rejection by total uncertainty). Additionally, the feature-importance experiment uses an ad-hoc rule—'rejecting 30% of instances with highest epistemic uncertainty'—to 'better capture true aleatoric uncertainty' (§3.1.2). This threshold is not justified or validated, and the appendix reports a significance test only on the remaining 70% without showing that the rejected instances are actually those for which explanations are unreliable. Please quantify the effect of rejection on explanation quality (e.g., error rate, explanation stability) as a function of the rejection threshold, and compare the EU-based rejection with a random rejection baseline.

minor comments (7)

[Title] The title in the full text reads 'A P ATH' — presumably a typo for 'A PATH'.
[§1] 'many attemps' should be 'many attempts'.
[Eq. (3)] In Eq. (3), the condition 'h(xt) != yc' is ambiguous because yc is not explicitly defined as a class label different from h(xt). Please define yc clearly (e.g., 'yc is a class in Y such that yc \neq h(xt)').
[Appendix C] The p-values in Appendix C are all near zero even for near-zero correlations (e.g., Ionos. with AUcen: -0.04, p=5.60e-7) because the test sets are large. Report the effect sizes with confidence intervals or standardized effect sizes, and note that statistical significance does not imply practical importance.
[Appendix C and D] The centroid-based method simplifies the method of Van Amersfoort et al. (2020) by ignoring the learned weight matrix and using Euclidean distance on the raw features. This should be stated as a simplification, and its implications for the interpretation of AUcen should be acknowledged.
[§2.2] The neighborhood N(xt) is defined as the ball of radius ϵ, and the paper says L is estimated with 30 samples drawn from this neighborhood. Please specify how the 30 samples are drawn (e.g., uniform in the ball, or in a shell) and whether the same samples are used for each test point.
[Table 2] The header 'B. C ANCER' contains a spurious space; please fix to 'B. Cancer'.

Circularity Check

1 steps flagged · score 5.0 of 10

The strongest counterfactual correlations are partly built into the shared K=7 neighborhood structure; other evidence remains independent, so circularity is partial.

other [Section 3.1.1 and Appendix C (Eqs. 3, 7, 13)]
"A K-Nearest Neighbors classifier is trained with K = 7 and default parameters from the scikit-learn library. Counterfactual explanations for each prediction are generated according to Equation (3), and the attainability of each counterfactual is computed using Equations (7). ... For this experiment, we arbitrarily chose a number K of neighbors equal to 7. Aleatoric uncertainty is computed according to the Discord at Equation (13)."

The counterfactual dissimilarity d in Eq. (7) is the Euclidean distance to the nearest training point whose label differs from the KNN prediction (Eq. (3)), found in the same training set used for the K=7 classifier. AUbel in Eq. (13) is the discord of the belief function built from the K=7 nearest neighbors of the same test point in that same training set. The main driver of discord is the presence of multiple classes among those seven neighbors: when a different-class point is close enough to serve as the counterfactual, it is also likely to fall inside the K=7 ball and raise AUbel; conversely, a class-homogeneous K=7 ball keeps discord low.

full rationale

The paper's central protocol has independent content: the feature-importance correlations in Table 2, the AUcen and AUens columns in Table 1, and the MNIST rejection analysis do not reduce to the counterfactual definition. The uncertainty formulas themselves are standard (Shaker & Hüllermeier 2020; Van Amersfoort et al. 2020; Denœux 1995; Klir & Ramer 1990) rather than defined in terms of the explanation metrics. Self-citations such as Hoarau et al. (2024, 2025) and Nguyen et al. (2022) point to methods whose equations are either given in the paper or traceable to external references, so they are not load-bearing in a circular way. The one significant circularity concern is the AUbel counterfactual correlation: both the aleatoric uncertainty estimate and the counterfactual dissimilarity are constructed from the same K=7 nearest-neighbor geometry, so the headline negative correlations are substantially entangled with that shared construction. Because the paper explicitly privileges the local AUbel estimator on the strength of these correlations, this is a partial circularity rather than a merely cosmetic issue, but it does not invalidate the whole framework.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The framework pulls three UQ decompositions, two explanation methods, and robustness metrics from prior work. Its own contribution is the selection rule and the empirical correlations. The free parameters are mostly neighbor counts, bandwidths, and thresholds chosen by hand; no parameters are fitted to maximize the reported correlations, but rejecting 30% of instances is ad hoc and changes some results. The main axioms are that uncertainty decomposition is faithful, that distance and Lipschitz metrics proxy user-relevant robustness, and that rank correlations suffice to validate the protocol.

free parameters (9)

Number of neighbors K for KNN and Evidential K-NN = 7
Used for the classifier, counterfactual generation, and evidential uncertainty; chosen by hand in Appendix C and not fitted. It ties counterfactual distance and AUbel to the same neighborhood.
Lipschitz neighborhood radius epsilon = 0.1
Taken from Alvarez-Melis and Jaakkola (2018) in Eq (4)-(5); controls the SHAP robustness measure.
Number of perturbation samples for Lipschitz estimate = 30
From the same prior work; a small sample may make the local Lipschitz constant noisy.
Random forest size and depth for ensemble uncertainty = 100 estimators, max depth 4
Default-ish hyperparameters in Appendix C-D; AUens results are the weakest and most inconsistent, suggesting sensitivity to this choice.
Centroid length scale sigma = 1
RBF scale in Eq (10) is set ad hoc to 1 in all experiments.
Epistemic rejection quantile = 30%
The protocol rejects the top 30% of epistemically uncertain instances in feature-importance experiments; chosen by hand, and it changes the Wine correlation sign from -0.31 to +0.52.
MNIST training subset size = 2000
Deliberately small to create epistemically uncertain LeNet predictions; this is a design choice, not a deployment configuration.
Train/test split ratio = 70/30
Standard but arbitrary; results may vary with the split.
Number of experimental runs = 100 for counterfactuals, 5 for SHAP
Averaging schedule chosen by computational cost; the SHAP correlations rest on fewer runs.

assumptions (5)

domain assumption The three uncertainty estimators (evidential K-NN, centroid RBF, random forest ensembles) correctly separate aleatoric from epistemic uncertainty for real datasets.
Invoked throughout Section 2.3; there is no ground-truth decomposition on real data, so the disentanglement is assumed.
domain assumption Counterfactual distance (Eq 7) and local Lipschitz continuity (Eq 4) are appropriate proxies for explanation attainability and robustness.
The paper equates lower counterfactual distance with attainability and lower Lipschitz constant with robustness; no user study or task metric validates these proxies.
domain assumption Significant Spearman correlations between uncertainty and explanation metrics are sufficient to establish the value of the selection protocol.
The protocol is never evaluated end-to-end, such as accuracy of rejected instances, user preference, or explanation quality; observed correlations are treated as confirmation.
ad hoc to paper Rejecting the top 30% of epistemically uncertain instances yields a better estimate of true aleatoric uncertainty.
This assumption justifies the post-hoc rejection in Section 3.1.2; there is no independent validation that remaining instances better estimate AU.
domain assumption High aleatoric uncertainty means dense class-overlapping regions and high epistemic uncertainty means sparse regions.
This semantic link from the UQ literature underpins the routing rule in Figure 2; the paper does not verify it on real data beyond indirect correlations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust Explanations Through Uncertainty Decomposition: A Path to Trustworthier AI." pith.science (2026). https://pith.science/paper/FJE4TCFQ

@misc{pith2026250712913,
  author       = {Pith},
  title        = {Pith review of: Robust Explanations Through Uncertainty Decomposition: A Path to Trustworthier AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJE4TCFQ}},
  note         = {Machine review of arXiv:2507.12913}
}

read the original abstract

Recent advancements in machine learning have emphasized the need for transparency in model predictions, particularly as interpretability diminishes when using increasingly complex architectures. In this paper, we propose leveraging prediction uncertainty as a complementary approach to classical explainability methods. Specifically, we distinguish between aleatoric (data-related) and epistemic (model-related) uncertainty to guide the selection of appropriate explanations. Epistemic uncertainty serves as a rejection criterion for unreliable explanations and, in itself, provides insight into insufficient training (a new form of explanation). Aleatoric uncertainty informs the choice between feature-importance explanations and counterfactual explanations. This leverages a framework of explainability methods driven by uncertainty quantification and disentanglement. Our experiments demonstrate the impact of this uncertainty-aware approach on the robustness and attainability of explanations in both traditional machine learning and deep learning scenarios.

Figures

Figures reproduced from arXiv: 2507.12913 by the authors.

**Figure 1.** Uncertainty decomposition, with aleatoric uncertainty on the left (inherent difficulty of the classification [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Summary of the proposed method. may lead to a change in the model’s predictions. In such cases, one may need to use several features to construct informative feature-importance explanations. Such long and ambiguous explanations may challenge human cognitive ability, especially when the total number of features is large, which may be unavoidable in different applications. In contrast, by definition, counterfactual ex… view at source ↗

**Figure 3.** Aleatoric uncertainty vs. dissimilarity, leveraging a correlation between similarity and uncertainty. More aleatoric uncertainty leads to more similar counterfactual explanations. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: shows the relationship between aleatoric uncertainty and the un-robustness of SHAP explanations on four datasets. Samples with lower uncertainty tend to yield more stable SHAP explanations, reflecting higher robustness. This observed consistency enhances the reliabilit…

**Figure 5.** Figure 5: Rejection based on epistemic uncertainty on Glass dataset: 5a Rejection curve based on the epistemic [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Rejection based on epistemic uncertainty for MNIST dataset: 6a Rejection curve based on the epistemic un [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

[5]

aleatoric uncertainty In this experiment, we evaluate the robustness of feature-attribution explanations across ten datasets from the UCI repository (Dua & Graff, 2017)

D Feature-attribution robustness vs. aleatoric uncertainty In this experiment, we evaluate the robustness of feature-attribution explanations across ten datasets from the UCI repository (Dua & Graff, 2017). For each dataset, features are normalized and data is randomly split into training and test sets using a 70/30 ratio. Each experiment is repeated 5 ti...

work page 2017
[6]

A random forest with 100 estimators and a maximum depth of 4 is used

The experiment is also conducted using the entropy-based ensemble method presented in (Shaker & Hüllermeier, 2020). A random forest with 100 estimators and a maximum depth of 4 is used. All other parameters are set to the default values of the scikit-learn library (Pedregosa et al., 2011). To compute aleatoric uncertainty, we used the entropy decompositio...

work page 2020
[7]

This result is highlighted in Table

Aleatoric uncertainty is computed accord- ing to the Discord at Equation (13), as proposed by the authors in (Hoarau et al., 2024). This result is highlighted in Table

work page 2024
[8]

and Gales, M

Malinin, A. and Gales, M. Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019,

work page 2019
[9]

For the case of feature-importance explanations, we used the Glass dataset from the UCI repository (Dua & Graff, 2017)

E Rejection of explanations based on epistemic uncertainty For this experiment, we illustrate the proposition of rejecting explanations based on epistemic uncertainty in both machine learning and deep learning contexts. For the case of feature-importance explanations, we used the Glass dataset from the UCI repository (Dua & Graff, 2017). After normalizati...

work page 2017
[10]

Sundararajan, M

URLhttps: //arxiv.org/abs/2204.06507. Sundararajan, M. and Najmi, A. The many shapley values for model explanation. In International conference on machine learning, pp. 9269–9278. PMLR,

arXiv
[11]

How to choose an explainability method? towards a methodical implementation of xai in practice

Vermeire, T., Laugel, T., Renard, X., Martens, D., and Detyniecki, M. How to choose an explainability method? towards a methodical implementation of xai in practice. In ECML-PKDD 2021 Workshop: XKDD,

work page 2021
[12]

REPEAT: Improving Uncertainty Estimation in Representation Learning Explainability

URL https://arxiv.org/abs/2412.08513. 13 A PREPRINT A Notations and Acronyms Notations/Acronyms Description Y Class (random) variable X q A feature (random) variable X Feature vector {X 1, . . . , XQ} X q Range of feature X q X Input space X 1 × . . .× XQ Y Output space h A classifier Dtrain Training dataset {(⃗ xn, yn) | 1 ≤ n ≤ N } ⃗ xn n-th training in...

work page Pith review arXiv

Show all 21 references

[13]

For each dataset, the feature space is normalized, and the data is randomly split into training and test sets using a 70/30 ratio

are used, varying in size and complexity. For each dataset, the feature space is normalized, and the data is randomly split into training and test sets using a 70/30 ratio. Each experiment is repeated 100 times. A K-Nearest Neighbors classifier is trained with K = 7 and defaul...

2011
[18]

Cancer -0.83 ≃ 0 √ Ecoli -0.53 ≃ 0 √ Glass -0.55 ≃ 0 √ Heart -0.79 ≃ 0 √ Ionos

15 A PREPRINT Dataset Correlation p-value Significance B. Cancer -0.83 ≃ 0 √ Ecoli -0.53 ≃ 0 √ Glass -0.55 ≃ 0 √ Heart -0.79 ≃ 0 √ Ionos. -0.61 ≃ 0 √ Iris -0.58 ≃ 0 √ Liver -0.82 ≃ 0 √ Park. -0.89 ≃ 0 √ Sonar -0.82 ≃ 0 √ Wine -0.45 1.42 × 10−275 √ Table 5: Spearman’s rank corr...

2020
[19]

This method is only applicable to binary classification tasks, results are reported in Table 12 only for two-class datasets

For the likelihood-based approach, aleatoric uncertainty is computed based on the plausibility function introduced by (Nguyen et al., 2022). This method is only applicable to binary classification tasks, results are reported in Table 12 only for two-class datasets. Finally, th...

2022
[20]

Parameters for the method follow the setup described in (Hoarau et al., 2024), with K = 7 neighbors

is also evaluated. Parameters for the method follow the setup described in (Hoarau et al., 2024), with K = 7 neighbors. Aleatoric uncertainty is assessed using both Discord (13) and Non-specificity (12) measures, as proposed by the authors. These results are presented in Table

2024
[1995]

The parameters for the method are the same as those used in the version presented by (Hoarau et al., 2024)

is also studied to estimate aleatoric uncer- tainty. The parameters for the method are the same as those used in the version presented by (Hoarau et al., 2024). For this experiment, we arbitrarily chose a number K of neighbors equal to

2024
[1998]

D., and Li, Y

Liu, W., Wang, X., Owens, J. D., and Li, Y . Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, volume 2020-December,

2020
[2017]

M., Erion, G

Lundberg, S. M., Erion, G. G., Chen, H., DeGrave, A. J., Prutkin, J. M., Nair, B. G., Katz, R., Himmelfarb, J., Bansal, N., and Lee, S.-I. Explainable ai for trees: From local explanations to global understanding.ArXiv, abs/1905.04610,

1905 arXiv
[2018]

org/abs/1806.08049

URL https://arxiv. org/abs/1806.08049. Audemard, G., Bellart, S., Bounia, L., Koriche, F., Lagniez, J., and Marquis, P. On preferred abductive explanations for decision trees and random forests. In Proc. of IJCAI’22, 2022a. 10 A PREPRINT Audemard, G., Bellart, S., Bounia, L., ...

arXiv
[2020]

and Rad, P

Das, A. and Rad, P. Opportunities and challenges in explainable artificial intelligence (xai): A survey. arXiv preprint arXiv:2006.11371,

2006 arXiv
[2022]

doi: 10.1016/j.ijhcs.2022. 102839. Jiang, J., Leofante, F., Rago, A., and Toni, F. Robust counterfactual explanations in machine learning: A survey. In International Joint Conference on Artificial Intelligence,

2022 doi
[2023]

Learning model-agnostic counterfactual explanations for tabular data

Pawelczyk, M., Broelemann, K., and Kasneci, G. Learning model-agnostic counterfactual explanations for tabular data. In Proceedings of The Web Conference 2020 , WWW ’20, pp. 3126–3132, New York, NY , USA,

2020
[2024]

SAFARI: Versatile and Efficient Evaluations for Robustness of In- terpretability

Huang, W., Zhao, X., Jin, G., and Huang, X. SAFARI: Versatile and Efficient Evaluations for Robustness of In- terpretability . In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pp. 1988–1998, Los Alamitos, CA, USA, October

2023
[2025]

URL https://arxiv.org/abs/2501.18268. Hora, S. C. Aleatory and epistemic uncertainty in probability elicitation with an example from hazardous waste management. Reliability Engineering & System Safety, 54(2):217–223,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [5]

aleatoric uncertainty In this experiment, we evaluate the robustness of feature-attribution explanations across ten datasets from the UCI repository (Dua & Graff, 2017)

D Feature-attribution robustness vs. aleatoric uncertainty In this experiment, we evaluate the robustness of feature-attribution explanations across ten datasets from the UCI repository (Dua & Graff, 2017). For each dataset, features are normalized and data is randomly split into training and test sets using a 70/30 ratio. Each experiment is repeated 5 ti...

work page 2017

[2] [6]

A random forest with 100 estimators and a maximum depth of 4 is used

The experiment is also conducted using the entropy-based ensemble method presented in (Shaker & Hüllermeier, 2020). A random forest with 100 estimators and a maximum depth of 4 is used. All other parameters are set to the default values of the scikit-learn library (Pedregosa et al., 2011). To compute aleatoric uncertainty, we used the entropy decompositio...

work page 2020

[3] [7]

This result is highlighted in Table

Aleatoric uncertainty is computed accord- ing to the Discord at Equation (13), as proposed by the authors in (Hoarau et al., 2024). This result is highlighted in Table

work page 2024

[4] [8]

and Gales, M

Malinin, A. and Gales, M. Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019,

work page 2019

[5] [9]

For the case of feature-importance explanations, we used the Glass dataset from the UCI repository (Dua & Graff, 2017)

E Rejection of explanations based on epistemic uncertainty For this experiment, we illustrate the proposition of rejecting explanations based on epistemic uncertainty in both machine learning and deep learning contexts. For the case of feature-importance explanations, we used the Glass dataset from the UCI repository (Dua & Graff, 2017). After normalizati...

work page 2017

[6] [10]

Sundararajan, M

URLhttps: //arxiv.org/abs/2204.06507. Sundararajan, M. and Najmi, A. The many shapley values for model explanation. In International conference on machine learning, pp. 9269–9278. PMLR,

arXiv

[7] [11]

How to choose an explainability method? towards a methodical implementation of xai in practice

Vermeire, T., Laugel, T., Renard, X., Martens, D., and Detyniecki, M. How to choose an explainability method? towards a methodical implementation of xai in practice. In ECML-PKDD 2021 Workshop: XKDD,

work page 2021

[8] [12]

REPEAT: Improving Uncertainty Estimation in Representation Learning Explainability

URL https://arxiv.org/abs/2412.08513. 13 A PREPRINT A Notations and Acronyms Notations/Acronyms Description Y Class (random) variable X q A feature (random) variable X Feature vector {X 1, . . . , XQ} X q Range of feature X q X Input space X 1 × . . .× XQ Y Output space h A classifier Dtrain Training dataset {(⃗ xn, yn) | 1 ≤ n ≤ N } ⃗ xn n-th training in...

work page Pith review arXiv